This repository provides supplementary material for the following paper:
Jobst, D., Möller, A., and Groß, J., 2023. D-Vine GAM Copula based Quantile Regression with Application to Ensemble Postprocessing. doi: 10.48550/arXiv.2309.05603.
The data needed for reproducing the results is publicly available:
Jobst, David, Möller, Annette, & Groß, Jürgen. (2023). Data set for the ensemble postprocessing of 2m surface temperature forecasts in Germany for 24 hours lead time (0.1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.8127734
For the data license see here.
- Source: ECMWF (European Centre for Medium-Range Weather Forecasts)
- Gridded forecasts: 50-member ensemble forecasts
- Time range: 2015-01-02 to 2020-12-31
- Forecast leadtime: 24 hours
- Forecast initialization time: 12 UTC
- Area: Germany
- Resolution: 0.25 degrees
- Meteorological variables
| Variable | Description |
|---|---|
| t2m | 2m surface temperature |
| d2m | 2m surface dewpoint temperature |
| p | Surface pressure |
| sr | Surface solar radiation |
| u10m | 10m surface u-wind speed component |
| v10m | 10m surface v-wind speed component |
| r2m | 2m surface relative humidity |
| tcc | Total cloud cover |
| ws10m | 10m surface wind speed |
| wg10m | 10m surface wind gust |
- Source: DWD Climate Data Center (German Weather Service)
- Observation data: Hourly observations of the target variable (2m surface temperature)
- Number of stations: 462
- ECMWF forecasts: Bilinearly interpolated to the SYNOP stations and reduced to its mean (variable_mean) and standard deviation (variable_sd)
- Metadata
| Variable | Description |
|---|---|
| obs | Observation of 2m surface temperature |
| id | Station ID |
| name | Station name |
| lon | Longitude of station |
| lat | Latitude of station |
| elev | Elevation of station |
| date | Date |
| doy | Day of the year |
| sin1 | Sine-transformed day of the year |
| cos1 | Cosine-transformed day of the year |
All models except of the D-vine copula quantile regression (DVQR) are estimated based on the static training data 2015-2019. For the DVQR model estimation a day-by-day sliding training window is applied which uses training data of 2020 as well. Finally, all models are evaluated in the whole year 2020.
- crch: Local Ensemble Model Output Statistics (EMOS) and its gradient-boosted extension (EMOS-GB)
- vinereg: Local D-vine copula based quantile regression (DVQR)
- gamvinereg: Local D-vine GAM copula based quantile regression (GAM-DVQR)
- eppverification: For the verification of the ensemble postprocessing models